monitord/

units.rs

//! # units module
//!
//! All main systemd unit statistics. Counts of types of units, unit states and
//! queued jobs. We also house service specific statistics and system unit states.

use std::collections::HashMap;
use std::str::FromStr;
use std::sync::Arc;
use std::time::Instant;
use std::time::SystemTime;
use std::time::UNIX_EPOCH;

use struct_field_names_as_array::FieldNamesAsArray;
use thiserror::Error;
use tokio::sync::RwLock;
use tokio::sync::Semaphore;
use tokio::task::JoinSet;
use tracing::debug;
use tracing::error;
use tracing::warn;
use tracing::Instrument;
use zbus::zvariant::ObjectPath;
use zbus::zvariant::OwnedObjectPath;

#[derive(Error, Debug)]
pub enum MonitordUnitsError {
    #[error("Units D-Bus error: {0}")]
    ZbusError(#[from] zbus::Error),
    #[error("Integer conversion error: {0}")]
    IntConversion(#[from] std::num::TryFromIntError),
    #[error("System time error: {0}")]
    SystemTimeError(#[from] std::time::SystemTimeError),
}

use crate::timer::TimerStats;
use crate::MachineStats;

// Re-export the enums and function from unit_constants for backwards compatibility
pub use crate::unit_constants::is_unit_unhealthy;
pub use crate::unit_constants::is_unit_unhealthy_for_service;
pub use crate::unit_constants::SystemdUnitActiveState;
pub use crate::unit_constants::SystemdUnitLoadState;
pub use crate::unit_constants::SYSTEMD_SERVICE_SUFFIX;
pub use crate::unit_constants::SYSTEMD_TIMER_SUFFIX;

/// Inner timing breakdown for the units collector D-Bus phases.
///
/// Helps locate which step of unit collection dominates wall time when the
/// `units` collector is the slowest one in `MonitordStats::collector_timings`.
#[derive(serde::Serialize, serde::Deserialize, Clone, Debug, Default, PartialEq)]
pub struct UnitsCollectionTimings {
    /// Time for the systemd ListUnits D-Bus call (one batched call returning all units).
    pub list_units_ms: f64,
    /// Time for filesystem unit file stats collection (runs concurrently with list_units).
    pub unit_files_ms: f64,
    /// Time spent in the per-unit parse loop, including any per-unit D-Bus calls
    /// (timer property fetches, state stats, service stats).
    pub per_unit_loop_ms: f64,
    /// Number of timer units whose properties were fetched via D-Bus this run.
    pub timer_dbus_fetches: u64,
    /// Number of unit state D-Bus fetches this run (when state_stats_time_in_state is enabled).
    pub state_dbus_fetches: u64,
    /// Number of per-service D-Bus property fetches this run.
    pub service_dbus_fetches: u64,
    /// Slowest units (by per-unit collection duration, descending) this run,
    /// truncated to `units.slowest_units_count`. Empty when disabled (count 0).
    pub slowest_units: Vec<(String, f64)>,
}

/// Unit file counts for a scope (root or user), broken down by unit type.
#[derive(serde::Serialize, serde::Deserialize, Clone, Debug, Default, PartialEq)]
pub struct UnitFilesScope {
    /// Generated unit files by type (e.g. "service" => 2, "mount" => 5)
    pub generated: HashMap<String, u64>,
    /// Transient unit files by type (e.g. "service" => 10, "scope" => 6)
    pub transient: HashMap<String, u64>,
}

/// Unit file statistics collected from the filesystem for root and user scopes.
#[derive(serde::Serialize, serde::Deserialize, Clone, Debug, Default, PartialEq)]
pub struct UnitFilesStats {
    pub root: UnitFilesScope,
    pub user: UnitFilesScope,
}

#[derive(
    serde::Serialize, serde::Deserialize, Clone, Debug, Default, FieldNamesAsArray, PartialEq,
)]

/// Aggregated systemd unit statistics: counts by type, load state, active state,
/// plus optional per-service and per-timer detailed metrics
pub struct SystemdUnitStats {
    /// Number of units in the "activating" state (in the process of being started)
    pub activating_units: u64,
    /// Number of units in the "active" state (currently started and running)
    pub active_units: u64,
    /// Number of automount units (on-demand filesystem mount points)
    pub automount_units: u64,
    /// Number of device units (kernel devices exposed to systemd by udev)
    pub device_units: u64,
    /// Number of units in the "failed" state (exited with error, crashed, or timed out)
    pub failed_units: u64,
    /// Number of units in the "inactive" state (not currently running)
    pub inactive_units: u64,
    /// Number of pending jobs queued in the systemd job scheduler
    pub jobs_queued: u64,
    /// Number of units whose unit file has been successfully loaded into memory
    pub loaded_units: u64,
    /// Number of units whose unit file is masked (symlinked to /dev/null, cannot be started)
    pub masked_units: u64,
    /// Number of mount units (filesystem mount points managed by systemd)
    pub mount_units: u64,
    /// Number of units whose unit file could not be found on disk
    pub not_found_units: u64,
    /// Number of path units (file/directory watch triggers)
    pub path_units: u64,
    /// Number of scope units (externally created process groups, e.g. user sessions)
    pub scope_units: u64,
    /// Number of service units (daemon/process lifecycle management)
    pub service_units: u64,
    /// Number of slice units (resource management groups in the cgroup hierarchy)
    pub slice_units: u64,
    /// Number of socket units (IPC/network socket activation endpoints)
    pub socket_units: u64,
    /// Number of target units (synchronization points for grouping units)
    pub target_units: u64,
    /// Number of timer units (calendar/monotonic scheduled triggers)
    pub timer_units: u64,
    /// Number of timer units with Persistent=yes (triggers missed runs after downtime)
    pub timer_persistent_units: u64,
    /// Number of timer units with RemainAfterElapse=yes (stays loaded after firing)
    pub timer_remain_after_elapse: u64,
    /// Total number of units known to systemd (all types, all states)
    pub total_units: u64,
    /// Unit file statistics from the filesystem (e.g. generator output counts)
    pub unit_files: UnitFilesStats,
    /// Per-service detailed metrics keyed by unit name (e.g. "sshd.service")
    pub service_stats: HashMap<String, ServiceStats>,
    /// Per-timer detailed metrics keyed by unit name (e.g. "logrotate.timer")
    pub timer_stats: HashMap<String, TimerStats>,
    /// Per-unit active/load state tracking keyed by unit name
    pub unit_states: HashMap<String, UnitStates>,
    /// Inner timing breakdown for this collector. Zero-valued before the first
    /// run completes. On the varlink path, `list_units_ms`/`per_unit_loop_ms`
    /// cover the varlink fetch, local parse, and the two D-Bus phases that have
    /// no varlink equivalent yet — the oneshot type lookups and the timer
    /// backfill — so `service_dbus_fetches` and `timer_dbus_fetches` may both be
    /// nonzero there. Only `state_dbus_fetches` stays zero on that path.
    pub collection_timings: UnitsCollectionTimings,
}

/// Per-service metrics from the org.freedesktop.systemd1.Service and Unit D-Bus interfaces.
/// Ref: <https://www.freedesktop.org/software/systemd/man/org.freedesktop.systemd1.html>
#[derive(
    serde::Serialize, serde::Deserialize, Clone, Debug, Default, Eq, FieldNamesAsArray, PartialEq,
)]
pub struct ServiceStats {
    /// Realtime timestamp (usec since epoch) when the unit most recently entered the active state
    pub active_enter_timestamp: u64,
    /// Realtime timestamp (usec since epoch) when the unit most recently left the active state
    pub active_exit_timestamp: u64,
    /// Total CPU time consumed by this service's cgroup in nanoseconds
    pub cpuusage_nsec: u64,
    /// Realtime timestamp (usec since epoch) when the unit most recently left the inactive state
    pub inactive_exit_timestamp: u64,
    /// Total bytes read from block I/O by this service's cgroup
    pub ioread_bytes: u64,
    /// Total number of block I/O read operations by this service's cgroup
    pub ioread_operations: u64,
    /// Memory available to the service (MemoryAvailable from cgroup), in bytes
    pub memory_available: u64,
    /// Current memory usage of the service's cgroup in bytes
    pub memory_current: u64,
    /// Number of times systemd has restarted this service (automatic restarts)
    pub nrestarts: u32,
    /// Current number of processes in this service's cgroup
    pub processes: u32,
    /// Configured restart delay for this service in microseconds (RestartUSec)
    pub restart_usec: u64,
    /// Realtime timestamp (usec since epoch) of the most recent state change of any kind
    pub state_change_timestamp: u64,
    /// errno-style exit status code from the main process (0 = success)
    pub status_errno: i32,
    /// Current number of tasks (threads) in this service's cgroup
    pub tasks_current: u64,
    /// Timeout in microseconds for the cleanup of resources after the service exits
    pub timeout_clean_usec: u64,
    /// Watchdog timeout in microseconds; the service must ping within this interval or be killed
    pub watchdog_usec: u64,
}

/// Per-unit state tracking combining active state, load state, and computed health.
/// Ref: <https://www.freedesktop.org/software/systemd/man/org.freedesktop.systemd1.html>
#[derive(
    serde::Serialize, serde::Deserialize, Clone, Debug, Default, Eq, FieldNamesAsArray, PartialEq,
)]
pub struct UnitStates {
    /// Current active state of the unit (active, inactive, failed, activating, deactivating, reloading)
    pub active_state: SystemdUnitActiveState,
    /// Current load state of the unit (loaded, error, masked, not_found)
    pub load_state: SystemdUnitLoadState,
    /// Computed health flag: true when a loaded unit is not active, or when load state is error/not_found.
    /// Masked units are never marked unhealthy since masking is an intentional admin action.
    /// Optional config can ignore inactive oneshot services.
    pub unhealthy: bool,
    /// Microseconds elapsed since the unit's most recent state change.
    /// None when time-in-state tracking is disabled in config (expensive D-Bus
    /// lookup per unit), or when no state change timestamp is known.
    pub time_in_state_usecs: Option<u64>,
}

// Declare state types
// Reference: https://www.freedesktop.org/software/systemd/man/org.freedesktop.systemd1.html
// SubState can be unit-type-specific so can't enum

#[derive(Debug)]
pub struct ListedUnit {
    pub name: String,                      // The primary unit name
    pub description: String,               // The human readable description
    pub load_state: String, // The load state (i.e. whether the unit file has been loaded successfully)
    pub active_state: String, // The active state (i.e. whether the unit is currently started or not)
    pub sub_state: String,    // The sub state (i.e. unit type more specific state)
    pub follow_unit: String, // A unit that is being followed in its state by this unit, if there is any, otherwise the empty string
    pub unit_object_path: OwnedObjectPath, // The unit object path
    pub job_id: u32, // If there is a job queued for the job unit, the numeric job id, 0 otherwise
    pub job_type: String, // The job type as string
    pub job_object_path: OwnedObjectPath, // The job object path
}
impl
    From<(
        String,
        String,
        String,
        String,
        String,
        String,
        OwnedObjectPath,
        u32,
        String,
        OwnedObjectPath,
    )> for ListedUnit
{
    fn from(
        tuple: (
            String,
            String,
            String,
            String,
            String,
            String,
            OwnedObjectPath,
            u32,
            String,
            OwnedObjectPath,
        ),
    ) -> Self {
        ListedUnit {
            name: tuple.0,
            description: tuple.1,
            load_state: tuple.2,
            active_state: tuple.3,
            sub_state: tuple.4,
            follow_unit: tuple.5,
            unit_object_path: tuple.6,
            job_id: tuple.7,
            job_type: tuple.8,
            job_object_path: tuple.9,
        }
    }
}

pub const SERVICE_FIELD_NAMES: &[&str] = &ServiceStats::FIELD_NAMES_AS_ARRAY;
pub const UNIT_FIELD_NAMES: &[&str] = &SystemdUnitStats::FIELD_NAMES_AS_ARRAY;
pub const UNIT_STATES_FIELD_NAMES: &[&str] = &UnitStates::FIELD_NAMES_AS_ARRAY;

/// Pull out selected systemd service statistics
///
/// The seven cgroup-derived fields (`cpuusage_nsec`, `ioread_bytes`,
/// `ioread_operations`, `memory_current`, `memory_available`, `processes`,
/// `tasks_current`) come from cgroupfs via `crate::cgroup` (#221), read in
/// parallel with the remaining D-Bus properties. `fs_root` prefixes the
/// cgroup mount — empty for the host, `/proc/<leader>/root` for containers.
/// `host_memory` is read once per collection cycle in `parse_unit_state` and
/// shared by every unit. A cgroup field that cgroupfs has no data for falls
/// back to its D-Bus property, which is also what covers cgroup v1 hosts
/// (no v2 files → all fallback, no new failure mode).
#[tracing::instrument(level = "debug", skip(connection, object_path))]
async fn parse_service(
    connection: &zbus::Connection,
    name: &str,
    object_path: &OwnedObjectPath,
    fs_root: &str,
    host_memory: Option<crate::cgroup::HostMemory>,
) -> Result<ServiceStats, MonitordUnitsError> {
    debug!("Parsing service {} stats", name);

    let sp = crate::dbus::zbus_service::ServiceProxy::builder(connection)
        .cache_properties(zbus::proxy::CacheProperties::No)
        .path(object_path.clone())?
        .build()
        .await?;
    let up = crate::dbus::zbus_unit::UnitProxy::builder(connection)
        .cache_properties(zbus::proxy::CacheProperties::No)
        .path(object_path.clone())?
        .build()
        .await?;

    // The cgroup path gates the filesystem read; the main and control PIDs
    // are folded into the process count the way `GetProcesses` does (both
    // can sit outside the cgroup).
    // Use tokio::join! without tokio::spawn to avoid per-task allocation overhead.
    // These all share the same D-Bus connection so spawn adds no parallelism benefit.
    let (control_group, main_pid, control_pid) =
        tokio::join!(sp.control_group(), sp.main_pid(), sp.control_pid());
    let (control_group, main_pid, control_pid) = (control_group?, main_pid?, control_pid?);
    let mut extra_pids = Vec::with_capacity(2);
    for pid in [main_pid, control_pid] {
        if pid != 0 {
            extra_pids.push(pid);
        }
    }

    // The filesystem read joins the remaining D-Bus properties in one batch,
    // so cgroupfs IO never serializes behind IPC. An empty cgroup path
    // (inactive unit) short-circuits inside `read_service_cgroup` and every
    // field below falls back, exactly like an unreadable cgroup.
    let cgroup_stats =
        crate::cgroup::read_service_cgroup(fs_root, &control_group, &extra_pids, host_memory);

    let (
        active_enter_timestamp,
        active_exit_timestamp,
        inactive_exit_timestamp,
        nrestarts,
        restart_usec,
        state_change_timestamp,
        status_errno,
        timeout_clean_usec,
        watchdog_usec,
        cgroup_stats,
    ) = tokio::join!(
        up.active_enter_timestamp(),
        up.active_exit_timestamp(),
        up.inactive_exit_timestamp(),
        sp.nrestarts(),
        sp.restart_usec(),
        up.state_change_timestamp(),
        sp.status_errno(),
        sp.timeout_clean_usec(),
        sp.watchdog_usec(),
        cgroup_stats,
    );

    // Per-field D-Bus fallback for whatever cgroupfs had no data for: only
    // missing fields pay for a fallback call, in the same single batch — so
    // the hot path (cgroup present, the common case for running services)
    // issues none, and a property that errors only fails its own field via
    // the `?` below rather than the whole service (matching how a missing
    // cgroup file behaves).
    let (
        cpuusage_nsec,
        ioread_bytes,
        ioread_operations,
        memory_current,
        memory_available,
        tasks_current,
    ) = tokio::join!(
        async {
            match cgroup_stats.cpu_usage_nsec {
                Some(value) => Ok(value),
                None => sp.cpuusage_nsec().await,
            }
        },
        async {
            match cgroup_stats.io_read_bytes {
                Some(value) => Ok(value),
                None => sp.ioread_bytes().await,
            }
        },
        async {
            match cgroup_stats.io_read_operations {
                Some(value) => Ok(value),
                None => sp.ioread_operations().await,
            }
        },
        async {
            match cgroup_stats.memory_current {
                Some(value) => Ok(value),
                None => sp.memory_current().await,
            }
        },
        async {
            match cgroup_stats.memory_available {
                Some(value) => Ok(value),
                None => sp.memory_available().await,
            }
        },
        async {
            match cgroup_stats.tasks_current {
                Some(value) => Ok(value),
                None => sp.tasks_current().await,
            }
        },
    );
    let (
        cpuusage_nsec,
        ioread_bytes,
        ioread_operations,
        memory_current,
        memory_available,
        tasks_current,
    ) = (
        cpuusage_nsec?,
        ioread_bytes?,
        ioread_operations?,
        memory_current?,
        memory_available?,
        tasks_current?,
    );

    Ok(ServiceStats {
        active_enter_timestamp: active_enter_timestamp?,
        active_exit_timestamp: active_exit_timestamp?,
        cpuusage_nsec,
        inactive_exit_timestamp: inactive_exit_timestamp?,
        ioread_bytes,
        ioread_operations,
        memory_current,
        memory_available,
        nrestarts: nrestarts?,
        processes: cgroup_stats.processes,
        restart_usec: restart_usec?,
        state_change_timestamp: state_change_timestamp?,
        status_errno: status_errno?,
        tasks_current,
        timeout_clean_usec: timeout_clean_usec?,
        watchdog_usec: watchdog_usec?,
    })
}

/// Compute microseconds in the current state from a state-change timestamp.
///
/// A zero timestamp means no state change was ever recorded (or the lookup
/// failed and was zeroed by the caller): time in state is unknown, so this
/// returns `None` rather than a bogus huge elapsed time. Shared by the D-Bus
/// path (`get_time_in_state`) and the varlink path so both agree.
pub(crate) fn compute_time_in_state(now_usec: u64, state_change_timestamp: u64) -> Option<u64> {
    if state_change_timestamp == 0 {
        return None;
    }
    Some(now_usec.saturating_sub(state_change_timestamp))
}

#[tracing::instrument(level = "debug", skip(connection))]
async fn get_time_in_state(
    connection: Option<&zbus::Connection>,
    unit: &ListedUnit,
) -> Result<Option<u64>, MonitordUnitsError> {
    match connection {
        Some(c) => {
            let up = crate::dbus::zbus_unit::UnitProxy::builder(c)
                .cache_properties(zbus::proxy::CacheProperties::No)
                .path(ObjectPath::from(unit.unit_object_path.clone()))?
                .build()
                .await?;
            let now: u64 = SystemTime::now().duration_since(UNIX_EPOCH)?.as_secs() * 1_000_000;
            let state_change_timestamp = match up.state_change_timestamp().await {
                Ok(sct) => sct,
                Err(err) => {
                    error!(
                        "Unable to get state_change_timestamp for {} - Setting to 0: {:?}",
                        &unit.name, err,
                    );
                    0
                }
            };
            Ok(compute_time_in_state(now, state_change_timestamp))
        }
        None => {
            error!("No zbus connection passed, but time_in_state_usecs enabled");
            Ok(None)
        }
    }
}

/// Parse state of a unit into a `UnitStates` entry for the caller to merge.
///
/// Returns `(did_dbus_fetch, entry)`: `did_dbus_fetch` is true when an actual
/// time-in-state D-Bus fetch was performed, so callers can keep
/// `state_dbus_fetches` honest. Allowlist/blocklist short-circuits return
/// `(false, None)` — the unit is simply not tracked in `unit_states`.
#[tracing::instrument(level = "debug", skip(config, connection))]
pub async fn parse_state(
    unit: &ListedUnit,
    config: &crate::config::UnitsConfig,
    connection: Option<&zbus::Connection>,
) -> Result<(bool, Option<UnitStates>), MonitordUnitsError> {
    if config.state_stats_blocklist.contains(&unit.name) {
        debug!("Skipping state stats for {} due to blocklist", &unit.name);
        return Ok((false, None));
    }
    if !config.state_stats_allowlist.is_empty()
        && !config.state_stats_allowlist.contains(&unit.name)
    {
        return Ok((false, None));
    }
    let active_state = SystemdUnitActiveState::from_str(&unit.active_state)
        .unwrap_or(SystemdUnitActiveState::unknown);
    let load_state = SystemdUnitLoadState::from_str(&unit.load_state.replace('-', "_"))
        .unwrap_or(SystemdUnitLoadState::unknown);
    let mut is_oneshot_service = false;
    if config.ignore_inactive_oneshot_services
        && unit.name.ends_with(SYSTEMD_SERVICE_SUFFIX)
        && matches!(active_state, SystemdUnitActiveState::inactive)
        && matches!(load_state, SystemdUnitLoadState::loaded)
    {
        if let Some(conn) = connection {
            match is_oneshot_service_unit(conn, unit).await {
                Ok(is_oneshot) => is_oneshot_service = is_oneshot,
                Err(err) => warn!(
                    "Unable to get Service.Type for {} (assuming not oneshot): {:?}",
                    &unit.name, err
                ),
            }
        }
    }

    // Get the state_change_timestamp to determine time in usecs we've been in current state
    let mut time_in_state_usecs: Option<u64> = None;
    let mut did_dbus_fetch = false;
    if config.state_stats_time_in_state {
        time_in_state_usecs = get_time_in_state(connection, unit).await?;
        // get_time_in_state only issues a D-Bus call when connection is Some;
        // the None path logs an error and returns Ok(None) without calling out.
        did_dbus_fetch = connection.is_some();
    }

    let entry = UnitStates {
        active_state,
        load_state,
        unhealthy: is_unit_unhealthy_for_service(
            active_state,
            load_state,
            is_oneshot_service,
            config.ignore_inactive_oneshot_services,
        ),
        time_in_state_usecs,
    };
    Ok((did_dbus_fetch, Some(entry)))
}

#[tracing::instrument(level = "debug", skip(connection))]
async fn is_oneshot_service_unit(
    connection: &zbus::Connection,
    unit: &ListedUnit,
) -> Result<bool, MonitordUnitsError> {
    is_oneshot_service_at_path(connection, &unit.unit_object_path).await
}

/// Check whether a service unit is `Type=oneshot` via its D-Bus object path.
async fn is_oneshot_service_at_path(
    connection: &zbus::Connection,
    object_path: &OwnedObjectPath,
) -> Result<bool, MonitordUnitsError> {
    let sp = crate::dbus::zbus_service::ServiceProxy::builder(connection)
        .cache_properties(zbus::proxy::CacheProperties::No)
        .path(ObjectPath::from(object_path.clone()))?
        .build()
        .await?;
    Ok(sp.type_().await? == "oneshot")
}

/// Check whether a service unit is `Type=oneshot`, resolving the unit name to
/// a D-Bus object path first.
///
/// Used by the varlink units path, which knows unit names (from metric
/// objects) but never sees D-Bus object paths.
pub(crate) async fn is_oneshot_service_by_name(
    connection: &zbus::Connection,
    unit_name: &str,
) -> Result<bool, MonitordUnitsError> {
    let mp = crate::dbus::zbus_systemd::ManagerProxy::builder(connection)
        .cache_properties(zbus::proxy::CacheProperties::No)
        .build()
        .await?;
    let object_path = mp.get_unit(unit_name).await?;
    is_oneshot_service_at_path(connection, &object_path).await
}

/// Parse a unit and add to overall counts of state, type etc.
fn parse_unit(stats: &mut SystemdUnitStats, unit: &ListedUnit) {
    // Count unit type
    match unit.name.rsplit('.').next() {
        Some("automount") => stats.automount_units += 1,
        Some("device") => stats.device_units += 1,
        Some("mount") => stats.mount_units += 1,
        Some("path") => stats.path_units += 1,
        Some("scope") => stats.scope_units += 1,
        Some("service") => stats.service_units += 1,
        Some("slice") => stats.slice_units += 1,
        Some("socket") => stats.socket_units += 1,
        Some("target") => stats.target_units += 1,
        Some("timer") => stats.timer_units += 1,
        unknown => debug!("Found unhandled '{:?}' unit type", unknown),
    };
    // Count load state
    match unit.load_state.as_str() {
        "loaded" => stats.loaded_units += 1,
        "masked" => stats.masked_units += 1,
        "not-found" => stats.not_found_units += 1,
        _ => debug!("{} is not loaded. It's {}", unit.name, unit.load_state),
    };
    // Count unit status
    match unit.active_state.as_str() {
        "activating" => stats.activating_units += 1,
        "active" => stats.active_units += 1,
        "failed" => stats.failed_units += 1,
        "inactive" => stats.inactive_units += 1,
        unknown => debug!("Found unhandled '{}' unit state", unknown),
    };
    // Count jobs queued
    if unit.job_id != 0 {
        stats.jobs_queued += 1;
    }
}

const TRANSIENT_DIR: &str = "/run/systemd/transient";

async fn count_unit_files_by_type(path: &str) -> HashMap<String, u64> {
    let mut dir = match tokio::fs::read_dir(path).await {
        Ok(d) => d,
        Err(err) => {
            debug!("Unable to read {}: {:?}", path, err);
            return HashMap::new();
        }
    };
    let mut counts = HashMap::new();
    loop {
        match dir.next_entry().await {
            Ok(Some(entry)) => {
                let file_type = match entry.file_type().await {
                    Ok(ft) => ft,
                    Err(_) => continue,
                };
                if !file_type.is_file() {
                    continue;
                }
                let name = entry.file_name();
                let unit_type = name
                    .to_str()
                    .and_then(|n| n.rsplit('.').next())
                    .unwrap_or("unknown");
                *counts.entry(unit_type.to_string()).or_insert(0) += 1;
            }
            Ok(None) => break,
            Err(err) => {
                warn!("Error reading entry in {}: {:?}", path, err);
                continue;
            }
        }
    }
    counts
}

fn merge_counts(target: &mut HashMap<String, u64>, source: HashMap<String, u64>) {
    for (unit_type, count) in source {
        *target.entry(unit_type).or_insert(0) += count;
    }
}

/// Enumerate the per-user systemd transient directories under `{fs_root}/run/user`.
async fn enumerate_user_transient_dirs(fs_root: &str) -> Vec<String> {
    let user_dir = format!("{fs_root}/run/user");
    match tokio::fs::read_dir(&user_dir).await {
        Ok(mut entries) => {
            let mut dirs = Vec::new();
            loop {
                match entries.next_entry().await {
                    Ok(Some(entry)) => {
                        dirs.push(format!("{}/systemd/transient", entry.path().display()));
                    }
                    Ok(None) => break,
                    Err(err) => {
                        warn!("Error reading entry in {}: {:?}", user_dir, err);
                        continue;
                    }
                }
            }
            dirs
        }
        Err(err) => {
            debug!("Unable to read {}: {:?}", user_dir, err);
            Vec::new()
        }
    }
}

/// Collect unit file statistics from the filesystem.
/// `fs_root` is prepended to all paths — empty string for the host,
/// `/proc/<pid>/root` for containers.
///
/// All directory reads are issued in parallel: the three generator directories,
/// the root transient directory, and user-dir enumeration run concurrently in a
/// first batch; per-user transient reads run concurrently in a second batch.
pub async fn collect_unit_files_stats(fs_root: &str) -> UnitFilesStats {
    // Pre-bind formatted paths to extend their lifetime across the join.
    let gen_path = format!("{fs_root}/run/systemd/generator");
    let gen_early_path = format!("{fs_root}/run/systemd/generator.early");
    let gen_late_path = format!("{fs_root}/run/systemd/generator.late");
    let transient_path = format!("{fs_root}{TRANSIENT_DIR}");

    // First batch: fixed paths + user dir enumeration all in parallel.
    let (gen, gen_early, gen_late, root_transient, user_dirs) = tokio::join!(
        count_unit_files_by_type(&gen_path),
        count_unit_files_by_type(&gen_early_path),
        count_unit_files_by_type(&gen_late_path),
        count_unit_files_by_type(&transient_path),
        enumerate_user_transient_dirs(fs_root),
    );

    let mut root_generated = HashMap::new();
    merge_counts(&mut root_generated, gen);
    merge_counts(&mut root_generated, gen_early);
    merge_counts(&mut root_generated, gen_late);

    // Second batch: read every user transient directory in parallel.
    let user_transient_counts =
        futures_util::future::join_all(user_dirs.iter().map(|d| count_unit_files_by_type(d))).await;

    let mut user_transient = HashMap::new();
    for counts in user_transient_counts {
        merge_counts(&mut user_transient, counts);
    }

    UnitFilesStats {
        root: UnitFilesScope {
            generated: root_generated,
            transient: root_transient,
        },
        user: UnitFilesScope {
            generated: HashMap::new(),
            transient: user_transient,
        },
    }
}

/// Owned result of one unit's concurrent D-Bus work, merged into `SystemdUnitStats`
/// by the caller once the spawned task completes. Keeping this owned (rather than
/// mutating `SystemdUnitStats` from multiple concurrent tasks) avoids needing a
/// `Mutex`/`RwLock` around it.
#[derive(Default)]
struct PerUnitOutcome {
    unit_name: String,
    unit_states_entry: Option<UnitStates>,
    state_dbus_fetch: bool,
    service_stats_entry: Option<ServiceStats>,
    timer_stats_entry: Option<TimerStats>,
    duration_ms: f64,
}

/// Pull all units from dbus and count how system is setup and behaving
#[tracing::instrument(level = "debug", skip(config, connection))]
pub async fn parse_unit_state(
    config: &Arc<crate::config::Config>,
    connection: &zbus::Connection,
    fs_root: &str,
) -> Result<SystemdUnitStats, MonitordUnitsError> {
    if !config.units.state_stats_allowlist.is_empty() {
        debug!(
            "Using unit state allowlist: {:?}",
            config.units.state_stats_allowlist
        );
    }

    if !config.units.state_stats_blocklist.is_empty() {
        debug!(
            "Using unit state blocklist: {:?}",
            config.units.state_stats_blocklist,
        );
    }

    let mut stats = SystemdUnitStats::default();

    let p = crate::dbus::zbus_systemd::ManagerProxy::builder(connection)
        .cache_properties(zbus::proxy::CacheProperties::No)
        .build()
        .await?;

    // Run filesystem collection and D-Bus list_units in parallel, timing each independently.
    let (unit_files_result, units_result) = tokio::join!(
        async {
            let start = Instant::now();
            let files = if config.units.unit_files {
                collect_unit_files_stats(fs_root).await
            } else {
                UnitFilesStats::default()
            };
            (files, start.elapsed().as_secs_f64() * 1000.0)
        },
        async {
            let start = Instant::now();
            let units = p.list_units().await;
            (units, start.elapsed().as_secs_f64() * 1000.0)
        },
    );
    let (unit_files, unit_files_ms) = unit_files_result;
    let (units_result, list_units_ms) = units_result;
    stats.collection_timings.unit_files_ms = unit_files_ms;
    stats.collection_timings.list_units_ms = list_units_ms;
    stats.unit_files = unit_files;

    let units = units_result?;
    stats.total_units = units.len() as u64;

    let per_unit_loop_start = Instant::now();
    let mut state_dbus_fetches: u64 = 0;
    let mut service_dbus_fetches: u64 = 0;
    let mut timer_dbus_fetches: u64 = 0;

    // Cheap synchronous unit-type/state counting first, separate from the
    // concurrent D-Bus work below — no .await, so no reason to involve the
    // per-unit tasks in it.
    let listed_units: Vec<ListedUnit> = units.into_iter().map(ListedUnit::from).collect();
    for unit in &listed_units {
        parse_unit(&mut stats, unit);
    }

    // Bounded-concurrency D-Bus work per unit. A semaphore (rather than an
    // unbounded join) caps how many units are in flight at once: a burst of
    // simultaneous D-Bus calls could itself worsen host-level IPC contention
    // on hosts where per-call latency is already elevated, which is exactly
    // the failure mode this loop exists to avoid.
    let semaphore = Arc::new(Semaphore::new(
        config.units.per_unit_concurrency.max(1) as usize
    ));
    // Captured once, outside the loop: `JoinSet::spawn` runs each unit on a new
    // task, and tracing spans do not cross task boundaries automatically. Without
    // attaching this as the explicit parent below, every `unit_collect` span
    // becomes its own unrelated root trace instead of a child of this collector.
    let parent_span = tracing::Span::current();
    // Host-wide memory numbers for the `memory_available` fold, read once
    // per collection cycle and shared by every unit (rather than once per
    // service inside `parse_service`).
    let host_memory = crate::cgroup::read_host_memory().await;
    let mut join_set: JoinSet<PerUnitOutcome> = JoinSet::new();
    for unit in listed_units {
        let semaphore = Arc::clone(&semaphore);
        let config = Arc::clone(config);
        let connection = connection.clone();
        let parent_span = parent_span.clone();
        let fs_root = fs_root.to_string();
        join_set.spawn(async move {
            let _permit = semaphore
                .acquire()
                .await
                .expect("semaphore closed unexpectedly");
            let unit_collect_start = Instant::now();
            let span =
                tracing::debug_span!(parent: &parent_span, "unit_collect", unit = %unit.name);
            let mut outcome = async {
                let mut outcome = PerUnitOutcome {
                    unit_name: unit.name.clone(),
                    ..Default::default()
                };

                // Collect per unit state stats - ActiveState + LoadState.
                // Not collecting SubState (yet). A D-Bus error on one unit
                // is logged and skipped rather than aborting the whole
                // collection cycle for every other unit.
                if config.units.state_stats {
                    match parse_state(&unit, &config.units, Some(&connection)).await {
                        Ok((did_fetch, entry)) => {
                            outcome.state_dbus_fetch = did_fetch;
                            outcome.unit_states_entry = entry;
                        }
                        Err(err) => {
                            error!("Unable to get state for {}: {:?}", unit.name, err);
                        }
                    }
                }

                // Collect service stats
                if config.services.contains(&unit.name) {
                    debug!("Collecting service stats for {:?}", &unit);
                    match parse_service(
                        &connection,
                        &unit.name,
                        &unit.unit_object_path,
                        &fs_root,
                        host_memory,
                    )
                    .await
                    {
                        Ok(service_stats) => outcome.service_stats_entry = Some(service_stats),
                        Err(err) => error!(
                            "Unable to get service stats for {} {}: {:#?}",
                            &unit.name, &unit.unit_object_path, err
                        ),
                    }
                }

                // Collect timer stats
                if config.timers.enabled
                    && unit.name.ends_with(SYSTEMD_TIMER_SUFFIX)
                    && !config.timers.blocklist.contains(&unit.name)
                    && (config.timers.allowlist.is_empty()
                        || config.timers.allowlist.contains(&unit.name))
                {
                    match crate::timer::collect_timer_stats(&connection, &unit).await {
                        Ok(ts) => outcome.timer_stats_entry = Some(ts),
                        Err(err) => error!("Failed to get {} stats: {:#?}", &unit.name, err),
                    }
                }

                outcome
            }
            .instrument(span)
            .await;

            outcome.duration_ms = unit_collect_start.elapsed().as_secs_f64() * 1000.0;
            outcome
        });
    }

    let mut slowest_units: Vec<(String, f64)> = Vec::new();
    while let Some(res) = join_set.join_next().await {
        let outcome = match res {
            Ok(outcome) => outcome,
            Err(err) => {
                error!("Per-unit collection task failed to join: {:?}", err);
                continue;
            }
        };
        if let Some(entry) = outcome.unit_states_entry {
            stats.unit_states.insert(outcome.unit_name.clone(), entry);
        }
        if outcome.state_dbus_fetch {
            state_dbus_fetches += 1;
        }
        if let Some(service_stats) = outcome.service_stats_entry {
            stats
                .service_stats
                .insert(outcome.unit_name.clone(), service_stats);
            service_dbus_fetches += 1;
        }
        if let Some(ts) = outcome.timer_stats_entry {
            if ts.persistent {
                stats.timer_persistent_units += 1;
            }
            if ts.remain_after_elapse {
                stats.timer_remain_after_elapse += 1;
            }
            stats.timer_stats.insert(outcome.unit_name.clone(), ts);
            timer_dbus_fetches += 1;
        }
        if config.units.slowest_units_count > 0 {
            slowest_units.push((outcome.unit_name, outcome.duration_ms));
        }
    }

    if config.units.slowest_units_count > 0 {
        slowest_units.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
        slowest_units.truncate(config.units.slowest_units_count as usize);
        stats.collection_timings.slowest_units = slowest_units;
    }

    let per_unit_loop_elapsed = per_unit_loop_start.elapsed();
    stats.collection_timings.per_unit_loop_ms = per_unit_loop_elapsed.as_secs_f64() * 1000.0;
    stats.collection_timings.state_dbus_fetches = state_dbus_fetches;
    stats.collection_timings.service_dbus_fetches = service_dbus_fetches;
    stats.collection_timings.timer_dbus_fetches = timer_dbus_fetches;

    debug!("unit stats: {:?}", stats);
    Ok(stats)
}

/// Async wrapper that can update unit stats when passed a locked struct.
/// `fs_root` is prepended to filesystem paths for unit file stats —
/// empty string for the host, `/proc/<pid>/root` for containers.
///
/// The whole (potentially hundreds-of-D-Bus-calls) collection runs before the
/// write lock is taken, not while holding it: `locked_machine_stats` is
/// shared by every collector, so holding it across the entire per-unit loop
/// would block every other collector's own (often much quicker) write until
/// this collection finished.
pub async fn update_unit_stats(
    config: Arc<crate::config::Config>,
    connection: zbus::Connection,
    locked_machine_stats: Arc<RwLock<MachineStats>>,
    fs_root: String,
) -> anyhow::Result<()> {
    let units_stats = parse_unit_state(&config, &connection, &fs_root).await;
    let mut machine_stats = locked_machine_stats.write().await;
    match units_stats {
        Ok(units_stats) => machine_stats.units = units_stats,
        Err(err) => error!("units stats failed: {:?}", err),
    }
    Ok(())
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::collections::HashSet;
    use strum::IntoEnumIterator;

    fn get_unit_file() -> ListedUnit {
        ListedUnit {
            name: String::from("apport-autoreport.timer"),
            description: String::from(
                "Process error reports when automatic reporting is enabled (timer based)",
            ),
            load_state: String::from("loaded"),
            active_state: String::from("inactive"),
            sub_state: String::from("dead"),
            follow_unit: String::from(""),
            unit_object_path: ObjectPath::try_from(
                "/org/freedesktop/systemd1/unit/apport_2dautoreport_2etimer",
            )
            .expect("Unable to make an object path")
            .into(),
            job_id: 0,
            job_type: String::from(""),
            job_object_path: ObjectPath::try_from("/").unwrap().into(),
        }
    }

    #[tokio::test]
    async fn test_state_parse() -> Result<(), MonitordUnitsError> {
        let test_unit_name = String::from("apport-autoreport.timer");
        let expected_stats = SystemdUnitStats {
            activating_units: 0,
            active_units: 0,
            automount_units: 0,
            device_units: 0,
            failed_units: 0,
            inactive_units: 0,
            jobs_queued: 0,
            loaded_units: 0,
            masked_units: 0,
            mount_units: 0,
            not_found_units: 0,
            path_units: 0,
            scope_units: 0,
            service_units: 0,
            slice_units: 0,
            socket_units: 0,
            target_units: 0,
            timer_units: 0,
            timer_persistent_units: 0,
            timer_remain_after_elapse: 0,
            total_units: 0,
            unit_files: UnitFilesStats::default(),
            service_stats: HashMap::new(),
            timer_stats: HashMap::new(),
            unit_states: HashMap::from([(
                test_unit_name.clone(),
                UnitStates {
                    active_state: SystemdUnitActiveState::inactive,
                    load_state: SystemdUnitLoadState::loaded,
                    unhealthy: true,
                    time_in_state_usecs: None,
                },
            )]),
            collection_timings: UnitsCollectionTimings::default(),
        };
        let mut stats = SystemdUnitStats::default();
        let systemd_unit = get_unit_file();
        let mut config = crate::config::UnitsConfig::default();

        // Test no allow list or blocklist; with connection: None, parse_state
        // takes the no-op path inside get_time_in_state and returns false.
        let (did_fetch, entry) = parse_state(&systemd_unit, &config, None).await?;
        if let Some(entry) = entry {
            stats.unit_states.insert(systemd_unit.name.clone(), entry);
        }
        assert_eq!(expected_stats, stats);
        assert!(!did_fetch);

        // Create an allow list
        config.state_stats_allowlist = HashSet::from([test_unit_name.clone()]);

        // test no blocklist and only allow list - Should equal the same as no lists above
        let mut allowlist_stats = SystemdUnitStats::default();
        let (did_fetch, entry) = parse_state(&systemd_unit, &config, None).await?;
        if let Some(entry) = entry {
            allowlist_stats
                .unit_states
                .insert(systemd_unit.name.clone(), entry);
        }
        assert_eq!(expected_stats, allowlist_stats);
        assert!(!did_fetch);

        // Now add a blocklist
        config.state_stats_blocklist = HashSet::from([test_unit_name]);

        // test blocklist with allow list (show it's preferred)
        let mut blocklist_stats = SystemdUnitStats::default();
        let expected_blocklist_stats = SystemdUnitStats::default();
        let (did_fetch, entry) = parse_state(&systemd_unit, &config, None).await?;
        if let Some(entry) = entry {
            blocklist_stats
                .unit_states
                .insert(systemd_unit.name.clone(), entry);
        }
        assert_eq!(expected_blocklist_stats, blocklist_stats);
        // Blocklist short-circuit must NOT count as a D-Bus fetch.
        assert!(!did_fetch);
        Ok(())
    }

    #[test]
    fn test_compute_time_in_state() {
        // Normal case: elapsed since the recorded change.
        assert_eq!(
            compute_time_in_state(1_700_000_010_000_000, 1_700_000_000_000_000),
            Some(10_000_000)
        );
        // Zero timestamp (never changed, or failed lookup zeroed by the
        // caller): unknown rather than a bogus huge elapsed time.
        assert_eq!(compute_time_in_state(1_700_000_010_000_000, 0), None);
        // Future timestamp (clock skew): saturates to 0 instead of underflowing.
        assert_eq!(compute_time_in_state(100, 200), Some(0));
    }

    #[test]
    fn test_unit_parse() {
        let expected_stats = SystemdUnitStats {
            activating_units: 0,
            active_units: 0,
            automount_units: 0,
            device_units: 0,
            failed_units: 0,
            inactive_units: 1,
            jobs_queued: 0,
            loaded_units: 1,
            masked_units: 0,
            mount_units: 0,
            not_found_units: 0,
            path_units: 0,
            scope_units: 0,
            service_units: 0,
            slice_units: 0,
            socket_units: 0,
            target_units: 0,
            timer_units: 1,
            timer_persistent_units: 0,
            timer_remain_after_elapse: 0,
            total_units: 0,
            unit_files: UnitFilesStats::default(),
            service_stats: HashMap::new(),
            timer_stats: HashMap::new(),
            unit_states: HashMap::new(),
            collection_timings: UnitsCollectionTimings::default(),
        };
        let mut stats = SystemdUnitStats::default();
        let systemd_unit = get_unit_file();
        parse_unit(&mut stats, &systemd_unit);
        assert_eq!(expected_stats, stats);
    }

    #[test]
    fn test_unit_parse_activating() {
        let mut activating_unit = get_unit_file();
        activating_unit.active_state = String::from("activating");
        let mut stats = SystemdUnitStats::default();
        parse_unit(&mut stats, &activating_unit);
        assert_eq!(stats.activating_units, 1);
        assert_eq!(stats.active_units, 0);
        assert_eq!(stats.inactive_units, 0);
    }

    #[test]
    fn test_iterators() {
        assert!(SystemdUnitActiveState::iter().collect::<Vec<_>>().len() > 0);
        assert!(SystemdUnitLoadState::iter().collect::<Vec<_>>().len() > 0);
    }

    #[tokio::test]
    async fn test_count_unit_files_by_type() {
        let dir = tempfile::tempdir().expect("Unable to create temp dir");
        let path = dir.path();

        std::fs::write(path.join("sshd.service"), "").unwrap();
        std::fs::write(path.join("nginx.service"), "").unwrap();
        std::fs::write(path.join("boot.mount"), "").unwrap();
        std::fs::write(path.join("swap.swap"), "").unwrap();
        std::fs::create_dir(path.join("multi-user.target.wants")).unwrap();

        let counts = count_unit_files_by_type(path.to_str().unwrap()).await;
        assert_eq!(counts.get("service"), Some(&2));
        assert_eq!(counts.get("mount"), Some(&1));
        assert_eq!(counts.get("swap"), Some(&1));
        assert_eq!(counts.get("wants"), None);
        assert_eq!(counts.len(), 3);
    }

    #[tokio::test]
    async fn test_count_unit_files_by_type_nonexistent_dir() {
        let counts = count_unit_files_by_type("/nonexistent/path").await;
        assert!(counts.is_empty());
    }

    #[tokio::test]
    async fn test_collect_unit_files_stats_with_fs_root() {
        let root = tempfile::tempdir().expect("Unable to create temp dir");
        let root_path = root.path();

        let gen_dir = root_path.join("run/systemd/generator");
        std::fs::create_dir_all(&gen_dir).unwrap();
        std::fs::write(gen_dir.join("boot.mount"), "").unwrap();
        std::fs::write(gen_dir.join("swap.swap"), "").unwrap();

        let transient_dir = root_path.join("run/systemd/transient");
        std::fs::create_dir_all(&transient_dir).unwrap();
        std::fs::write(transient_dir.join("run-thing.service"), "").unwrap();

        let user_transient = root_path.join("run/user/1000/systemd/transient");
        std::fs::create_dir_all(&user_transient).unwrap();
        std::fs::write(user_transient.join("app-code.scope"), "").unwrap();
        std::fs::write(user_transient.join("app-term.scope"), "").unwrap();

        let stats = collect_unit_files_stats(root_path.to_str().unwrap()).await;
        assert_eq!(stats.root.generated.get("mount"), Some(&1));
        assert_eq!(stats.root.generated.get("swap"), Some(&1));
        assert_eq!(stats.root.transient.get("service"), Some(&1));
        assert_eq!(stats.user.transient.get("scope"), Some(&2));
        assert!(stats.user.generated.is_empty());
    }
}